Datasets:

Olyxee
/

FinIR-IntentBench

license: apache-2.0
language: en
tags:
  - finance
  - structured-generation
  - finir
task_categories:
  - text-generation
pretty_name: FinIR-IntentBench
configs:
  - config_name: default
    data_files:
      - split: core
        path: data/core.jsonl
      - split: stress
        path: data/stress.jsonl

Dataset Card: FinIR-IntentBench v1

Purpose

Paired {natural-language finance instruction, expected FinIR Intent} examples for evaluating any FinIR-Intent compiler (the rule-based baseline, a fine-tuned model, or an LLM-backed one) against the canonical FinIR Intent Contract (schema 1.0, owned by the core finir package — finir.intent.json_schema() / schemas/finir-intent-v1.schema.json, runtime >=0.1.0,<0.2.0).

Size and splits

183 examples total.
core (143) — in-distribution phrasing the deterministic baseline is built to support.
stress (40) — a held-out subset of harder paraphrases (unlisted verbs, fractions, magnitude suffixes, idioms, out-of-domain phrasing) authored independently of the baseline's rules. The baseline was deliberately not tuned to these; they exist so evaluation reports a real coverage gap rather than a score the parser was fitted to. See the split methodology below and ../MODEL_CARD.md.

The difficulty field on every row records the split.

category	total	core	stress	covers
`valid_simple`	118	90	28	one operation (`relative_change` / `set` / `absolute_change`) across every supported target
`ambiguous`	21	16	5	vague quantities, no target, conflicting duplicate targets, "set to N%" on a non-percentage target
`unsupported`	17	13	4	clear but not a FinIR model mutation (acquisitions, IPOs/going public, layoffs, litigation, hiring, bankruptcy, spin-offs)
`multi_operation`	10	9	1	multiple simultaneous operations in one instruction
`invalid`	7	5	2	structurally valid, semantically wrong at execution (currency or unit mismatch)
`range`	6	6	0	a `range` batch sweep
`scenario`	4	4	0	named `scenarios`, each with simultaneous operations

Format

One JSON object per line in examples/intentbench_v1.jsonl:

{"id": "...", "category": "...", "difficulty": "core|stress", "text": "...",
 "expected_intent": { ...canonical FinIR Intent envelope... }}

9 examples (id prefixed canon_) instead carry a fixture field naming a file in the core repo's tests/fixtures/intents/ — the exact fixtures tests/test_intent_contract.py validates and executes on the runtime side. The repo copy references them by filename (rather than re-typing their JSON) so the benchmark and the runtime can never drift apart on those cases; the Hugging Face export (../release/huggingface/finir-intentbench/) resolves them inline so the dataset loads with no dependency on the GitHub repo (see "Loading" below).

Seven rows whose expected_intent is structurally "valid" but semantically wrong for the reference model carry "execution_expectation": "semantic_reject" — the evaluation checks the runtime correctly raises finir.intent.IntentValidationError, proving the NL layer transcribes faithfully and lets the runtime catch the error.

Loading (Hugging Face Datasets)

The Hugging Face export ships fully inlined JSONL (no fixture references; the expected_intent is a JSON string to keep a stable, flat schema), plus per-split files:

from datasets import load_dataset

ds = load_dataset("Olyxee/FinIR-IntentBench")  # 'core' and 'stress' splits
core = load_dataset("Olyxee/FinIR-IntentBench", split="core")

import json

row = core[0]
expected = json.loads(row["expected_intent"])  # the canonical envelope

No clone of the FinIR GitHub repo is required to use the dataset.

Generation methodology

Fully synthetic. No real company data, no data from any real financial system. All values are illustrative and generic.
Ground-truth expected intents are derived from a structured spec (target, operation, direction, magnitude) — not by running the baseline parser — so the benchmark measures the parser rather than being defined by it (../ generator in the workstream; see ../MODEL_CARD.md).
The 9 canon_* rows reuse the core repo's shared fixtures verbatim.
Human-reviewed: every row was reviewed for a correct canonical envelope and correct split label; the core subset is additionally gated by a test asserting the baseline reaches 100% status accuracy on it, and every expected intent is checked against the canonical JSON Schema in CI.

Split methodology (anti-leakage)

core is phrasing inside the documented rule set; stress is held-out paraphrases authored to fall outside it. A benchmark hand-tuned until the parser passes every case measures nothing — so the baseline is frozen against the stress subset. On the current baseline the stress subset scores well below core (status 0.70 vs 1.00, value 0.58 vs 1.00), which is the intended, honest signal. When a stress case is genuinely fixed in the rule set, it may move to core with regression coverage; new adversarial paraphrases are added to stress to keep the gap measurable.

What is not in this dataset

No period/time-scope field (forbidden by the v1.0 contract).
No private company, customer, or personal data — entirely synthetic.
No canonical alias ontology: targets are the raw FinIR model-input names; synonym → target mapping is the compiler's job, not the dataset's or the contract's.

Versioning

Filename-versioned (intentbench_v1.jsonl). A breaking change to the FinIR Intent Contract (a schema_version major bump) requires a new IntentBench major version, per docs/intent-contract.md#versioning-policy. Additive cases can be appended without a version bump.

License & attribution

Apache-2.0, matching the core FinIR repository. FinIR-IntentBench was contributed by Alisha Fatima (@AlishaFatima16) as part of the FinIR-Intent Hugging Face workstream; the canonical contract it targets is maintained by Olyxee.

Files

data/intentbench_v1.jsonl — all 183 examples (each expected_intent is a JSON string).
data/core.jsonl — 143 core examples.
data/stress.jsonl — 40 held-out stress examples.